Article
A River Basin over the Course of Time:
Multi-Temporal Analyses of Land Surface Dynamics
in the Yellow River Basin (China) Based on Medium
Resolution Remote Sensing Data
Christian Wohlfart 1, *, Gaohuan Liu 2 , Chong Huang 2,3 and Claudia Kuenzer 4
1
2
3
4
*
Company for Remote Sensing and Environmental Research (SLU), Kohlsteiner Strasse 5,
81243 Munich, Germany
State Key Laboratory of Resources and Environmental Information System, Institute of Geographic
Sciences and Natural Resources Research (IGSNRR), Chinese Academy of Sciences (CAS),
11A Da Tun Road, An Wai, Beijing 100101, China; [email protected] (G.L.); [email protected] (C.H.)
Jiangsu Center for Collaborative Innovation in Geographic Information Resource Development and
Application, Nanjing, Jiangsu 210023, China
German Aerospace Center (DLR), German Remote Sensing Data Center (DFD), Münchner Strasse 20,
82234 Oberpfaffenhofen, Germany; [email protected]
Correspondence: [email protected]; Tel.: +49-171-581-6950
Academic Editors: Magaly Koch and Prasad S. Thenkabail
Received: 8 December 2015; Accepted: 29 January 2016; Published: 26 February 2016
Abstract: The Yellow River Basin is one of China’s most densely-populated, fastest growing
and most dynamic regions, with abundant natural resources and intense agricultural production.
Major land policies have recently resulted in remarkable landscape modifications throughout
the basin. The availability of precise regional land cover change information is crucial to
better understand the prevailing dynamics and underlying factors influencing the current
processes in such a complex system and can additionally serve as a valuable component for
modeling and decision making. Such comprehensive and detailed information is lacking for the
Yellow River Basin so far. In this study, we derived land cover characteristics and dynamics from
the complete last decade based on optical high-temporal MODIS Normalized Differenced Vegetation
Index (NDVI) time series for the whole Yellow River Basin. After filtering and smoothing for
noise reduction with the use of the adaptive Savitzky–Golay filter, the processed time series was
used to derive a large variety of phenological and annual metrics. The final classifications for
the basin (2003 and 2013) were based on a random forest classifier, trained by reference samples
from very high-resolution imagery. The accuracy assessment for all 18 thematic classes, which
was based on a 30% reference data split, yielded an overall accuracy of 87% and 84% for 2003
and 2013, respectively. Major land cover and land use changes during the last decade have
occurred on the Loess Plateau, where land and conservation reforms triggered large-scale recovery
of grassland and shrubland habitat that had been previously covered by agriculture or sparse
vegetation. Agricultural encroachment and urban area expansion are other processes influencing the
dynamics in the basin. The necessity for regionally-adapted land cover maps becomes obvious when
our land cover products are compared to existing global products, where thematic accuracy remains
low, particularly in a heterogeneous landscape, such as the Yellow River Basin. The basin-wide
novel land cover and land use products of the Yellow River Basin hold a large potential for climate,
hydrology and biodiversity modelers, as well as river basin and regional governmental authorities
and will be shared upon request.
Keywords: land cover change; phenology; MODIS; random forest; time series analysis;
Yellow River Basin
Remote Sens. 2016, 8, 186; doi:10.3390/rs8030186
www.mdpi.com/journal/remotesensing
Remote Sens. 2016, 8, 186
2 of 25
1. Introduction
River basins provide a variety of ecosystem services and goods and, thus, have led to the dawn
of many ancient civilizations. Amongst others, they offer fertile soils, reliable water resources for
drinking and irrigation and an easy way of transportation. For much of China, the Yellow River
(Chinese: Huang He) is designated as the “cradle of civilization”. With early settlements along
the river banks, humans have shaped the river’s landscape and hydro-physical characteristics for
centuries. Today, the basin provides the livelihood for almost 190 million people and plays a vital
role in China’s continued economic development based on its abundant natural resources, including
rare ores, coal, oil and gas, which make it one of China’s most dynamic regions. Currently, the
basin’s provinces are attributed with annual GDP growth rates exceeding 10% [1]. As one of
China’s “breadbaskets”, the region contains approximately 13 million ha of arable land, in which
approximately 10 million tons of grain (in 2013) are produced, equivalent to nearly 20% of the
domestic production [1–4]. Hence, the high demand for and intense use of the prevailing water and
land resources have led to profound land use activities in the last few decades with adverse ecological,
economic and social repercussions [5–8]. A river basin requires holistic management perspectives and
reliable, consistent, multi-temporal information, and basin-wide land cover data are essential and can
serve as baseline information for the acting stakeholders to support sustainable regional management
applications and planning decisions.
Despite the urgent need to better understand the pressing land use dynamics in the
Yellow River Basin, there is a surprising lack of comprehensive land cover information for the entire
basin. In recent years, merely local land cover studies covering only small parts of the basin have been
conducted, mainly focusing on the source region of the Yellow River [9,10], the highly erosive Loess
Plateau in the middle reaches [11–13] or the fast-developing delta area [7,14]. To our knowledge,
only two basin-wide products published by Li et al. [15] and Wang et al. [16] exist. The former
study produced a thematic detailed mono-temporal land cover dataset for 1994, based on coarse
resolution AVHRR (Advanced Very High Resolution Radiometer) imagery. The latter work is a
multi-temporal change detection analysis between 1990 and 2000, which is derived from Landsat
Thematic Mapper (TM) data. Both products are outdated and not applicable for analyzing the current
and pressing dynamics in this region. Very often, the well-known global and continental datasets,
such as the Global Land Cover 2000 product (GLC 2000) [17], GlobCover [18], MODIS land cover
datasets [19] or the ESA Climate Change Initiative (CCI) Land Cover (CCI-LC) product [20], are often
used to answer regional research questions, and they serve as a basis for many land use management
plans and decisions. They have been derived from coarse to medium resolution Earth observation
satellites, such as AVHRR, MODIS (Moderate-resolution Imaging Spectroradiometer) or the Medium
Resolution Imaging Spectrometer (MERIS) instrument. Their ability for regionally-specific land
cover analysis and land cover dynamics is hampered, as they were designed for global applications
covering all vegetation and land cover types and, hence, are too generalized, particularly in
heterogeneous landscapes, such as the Yellow River Basin [21–26].
Phenology-driven land cover classification schemes on different regional scales using
high-temporal information have advanced, and the use of phenological variables as classification
input features has proven to be a reliable classification input for large areas with heterogenic
environmental conditions [14,27–29]. The MODIS sensor onboard the Terra and Aqua platforms can
observe the entire Earth’s surface with near daily global coverage. This multi-temporal datasets allow
for the composition of high temporal time series that capture the signatures of the prevailing land
cover entities, minimizing the influence of cloud coverage. Using this high temporal information on,
e.g., vegetation indices, it is possible to effectively discriminate different land cover classes based on
their intra-annual temporal signature [30].
Remote Sens. 2016, 8, 186
3 of 25
In order to fill the existing research gap, this study addresses the urgent need for consistent and
multi-temporal land cover products for the Yellow River Basin. We present a unique land cover
analysis at a spatial resolution of 250 m for the years 2003 and 2013 and additionally evaluated
the changes that occurred in this very dynamic region. The classification takes the advances
of phenology-based metrics, derived from high temporal resolution MODIS NDVI (Normalized
Differenced Vegetation Index) time series. This product is then compared to the previously-described
global land cover products. For this purpose, 1380 MODIS scenes of five MODIS tiles for six years had
to be processed. The generated consistent bi-temporal land cover data for the last decade are urgently
required and needed to improve our understanding about the sweeping dynamics in the basin.
2. Study Area
The Yellow River Basin (32◦ N–42◦ N/96◦ E–119◦ E) is a major river watershed in northern China,
and as the nation’s second largest basin, drains a total area of approximately 750,000 km2 [32].
Originating in the Qinghai-Tibet Plateau in western China, the river flows 5450 km, crossing nine
Chinese provinces and emptying into the Bohai Sea, where the Yellow River Delta is formed
(Figure 1). The climatic conditions are influenced by the continental monsoon circulation system.
Mean annual precipitation throughout the basin equals around 500 mm, but is unevenly distributed
across space and time [31]. The high altitude Qinghai-Tibet Plateau in the upper section receives
around 300 mm per year, and average temperatures remain low (1–4 ◦ C). Going further north, the
area becomes more arid, with only 150 mm of rainfall annually. The lower and middle parts of
the basin experience humid and sub-humid conditions with annual rainfall sums reaching up to
750 mm and temperatures ranging from 12–14 ◦ C. A large part of the annual precipitation falls
during the summer months (July and August), frequently in the form of heavy and intense summer
storms [33]. These events are mainly responsible for the high soil erosion rates and the large quantities
of sediment reaching the Yellow River and its tributaries [34]. Following the climatic gradient, the
basin encompasses a wide range of different ecoregions [35]: the vast high-alpine southeast Tibet
shrublands and meadows extend to the west, where nomadic pastoralism is still widespread. The
Gobi Desert and the arid Ordos Plateau steppe, located in the north, are poorly suited for agricultural
use due to infertile soils and scarce water resources. Further to the east lies the highly erosive
Central Loess Plateau, with dissected and erosive loessal hills, and the flat Huang He (or North
China) Plain, both of which are used intensively for agricultural production. The basin’s landscape
and river characteristics have been greatly affected by strong human activities, such as agriculture,
particularly irrigation, the extraction of natural resources, industrialization and dam/infrastructure
construction [6,16,36,37].
3. Material and Methods
The analytical framework of this study is based on high-temporal, medium-resolution Earth
observation data. The data processing and reference data collection procedure is summarized in
Figure 2. Seasonal and annual NDVI statistics based on MODIS time series served as input features
for the random forest classifier. The two final thematic products were validated using a 30% reference
data split. From the derived maps, we detected land cover changes and revealed local land surface
dynamics in the last decade, which were examined quantitatively and qualitatively. By comparing
local subregions, which are both complex and challenging, with high resolution Landsat imagery, we
demonstrate that regional land cover products outperform the available global products.
Remote Sens. 2016, 8, 186
4 of 25
Figure 1. The top figure depicts the Yellow River Basin and its major geographical units. The selected
diagrams show the different climatic patterns occurring in the basin. The underlying climate data
were derived from Hijmans et al. [31]. The bottom figure displays the provinces embedded in the
basin and the MODIS footprints covering the entire study area.
Remote Sens. 2016, 8, 186
5 of 25
Figure 2. Flowchart depicting the reference data collection steps (top) and the processing and
classification procedure (bottom) to derive the Yellow River Basin land cover product (YRB LC).
3.1. MODIS Data and Processing Steps
The MODIS surface reflectance product MOD09Q1 served as the backbone of the study and was
acquired for two time spans, from January 2002–December 2004 and from January 2012–December
2014 (Figure 2 (bottom)). Although we only focused on 2003 and 2013, it was necessary to process
three consecutive years for each time step, in order to guarantee the proper calculation of the
start/end of the season of the central year. In order to ensure full data coverage of the basin,
we selected a total of five MODIS tiles (h25v04, h25v05, h26v04, h26v05, h27v05) (Figure 1). The
MOD09Q1 8-day composite contains three data layers at a 250-m spatial resolution: MODIS Band
1 (red channel, 620–670 nm), Band 2 (near-infrared channel, 841–876 nm) and the MODIS Quality
Assurance (QA) layer [38]. A total of 1380 MODIS scenes were acquired and processed. For data
download and initial processing steps, we used the statistical programming environment R [39]
and the package MODIS, which incorporates automatic download and processing functions, such
as reprojection, mosaicking and subsetting [40]. For each scene, the native sinusoidal projection
was converted to the geographical latitude/longitude projection (WGS84 datum), keeping the
250-m spatial resolution. MODIS imagery is inherently influenced by atmospheric and geometric
interferences and noise; thus, we applied a quality assessment at the pixel level for each band and
used only the “highest” quality pixels provided by the QA layer. Low quality pixels were masked
out. With the cleaned bands, the Normalized Difference Vegetation Index (NDVI), which was based
on surface reflectance Bands 1 and 2, was calculated for the full 13-year time series with 46 time steps
each year. The interpolation of the gaps and additional noise reduction was achieved by applying
adaptive Savitzky–Golay (SG) filtering using TIMESAT software, which was specifically designed
for analyzing time series of satellite imagery [41]. The SG filter is a moving window filter that
performs a polynomial least-squares fit of a certain degree to all points in the window and replaces
the central point with the value of the polynomial [42]. In TIMESAT, the SG algorithm is implemented
in a newly-adapted form. Sometimes, the globally-defined window size does not account for rapid
increases or decreases in the NDVI trajectory. In these cases, it is necessary to locally decrease the
Remote Sens. 2016, 8, 186
6 of 25
window with a smaller one [41]. This filter type has the ability to follow rapid changes within the
time series. Within the Yellow River Basin, intensified agriculture with a two-season cropping cycle
is widespread, so we used a rather small window size of five as tested by Leinenkugel et al. [29] to
capture the fast changing phenology within a year.
3.2. Classification Features
The advantages of using phenological and annual metrics derived from high temporal and
medium spatial resolution remote sensing imagery as predictors for land cover classification have
been demonstrated in many broad-scale land cover studies, e.g., [30,43]. Every land cover class
consists of its individual temporal NDVI profile and is represented in the computed phenological
metrics for land cover discrimination, as shown in Figure 3. From the previously-filtered
and smoothed MODIS NDVI time series, we computed various seasonal phenology metrics,
characterizing the growing and biological cycle throughout the year. We derived eleven metrics
as defined and described by Jönsson and Eklundth [41] for 2003 and 2013. Furthermore, several
annual statistics were extracted for the NDVI and both spectral bands (RED and NIR), i.e., the median,
standard deviation or different percentiles (10, 25, 75, 90 percentiles) and their differences (75th–25th
percentile, 90th–10th percentile). Additionally, we added a digital elevation model (DEM) and slope
as additional predictors. In total, this results in 37 input features for the classifier (see Table 1).
Figure 3. Average temporal NDVI trajectories, derived from 8-day MOD09Q1 data, of all land
cover types occurring in the Yellow River Basin (green line) for 3 consecutive years (2012–2014); the
light-green shaded area represents the 10th and 90th percentile, respectively.
Remote Sens. 2016, 8, 186
7 of 25
Table 1. Phenological and annual metrics and their description. The latter were calculated for the
Vegetation Index and both spectral channels (RED, NIR).
Metrics
Description
Start of season
Peak value
Seasonal amplitude
Rate of increase at the beginning of the season
Rate of decrease at the end of the season
Large seasonal integral
Small seasonal integral
Time for which the left edge has increased to 40% of the seasonal amplitude measured
from the left minimum level
Time for which the right edge has decreased to 40% of the seasonal amplitude measured
from the right minimum level
Time from the start to the end of the season
Average of the left and right minimum values
Mean value of the times for which the left edge has increased to the 80% level and
the right edge has decreased to the 80% level
Largest value between the start and end of the season
Difference between the peak value and the base level
Ratio of the difference between the 20% and 80% levels and the corresponding time difference
Ratio of the difference between the right 20% and 80% levels and the corresponding time difference
Integral of all values between the start and the end of the season
Integral of all values from the start to the end of the season minus the base level
Median
Standard deviation
10th percentile
25th percentile
75th percentile
90th percentile
Diff 90th–10th percentile
Diff 75th–25th percentile
Median value derived from annual multi-temporal statistics
Standard deviation from annual multi-temporal statistics
10th percentile value from annual multi-temporal statistics
25th percentile value from annual multi-temporal statistics
75th percentile value from annual multi-temporal statistics
90the percentile value from annual multi-temporal statistics
Difference between 90th and 10thpercentile
Difference between 75 and 25 percentile
Annual metrics
Seasonal metrics
End of the season
Length of the season
Base level
Middle of the season
3.3. Reference Data Collection
For remote sensing-based land cover analysis, it is important to rely on a high quality reference
database for training and testing. A faulty and inappropriate sampling scheme can introduce much
error, influencing the reliability of the final classification. In September 2013, a two-week field
campaign was conducted in the Yellow River Basin. In the field, we collected suitable reference
samples of homogeneous areas. However, the collection of a valid amount of representative samples
to characterize the prevailing land cover for the entire study area was not possible in an in situ
field campaign throughout a region covering around 750,000 km2 . Therefore, we additionally used
high-resolution data from Landsat TM/OLI sensors and complementary very high-resolution (VHR)
imagery provided in virtual globe software, such as Google Earth and Mapworld, to collect a
comprehensive reference dataset for training and validation, as is often applied in large-scale and
medium-resolution remote sensing studies [26,44,45]. The VHR imagery available in our study region
stems mainly from Digital Globe’s Quickbird and GeoEye’s IKONOS satellites.
First, we defined adequate thematic land cover classes for the Yellow River Basin (see Table 2)
by interpreting various sources and databases, including existing region-specific land cover
descriptions, expert knowledge and VHR imagery in combination with geo-referenced photographs.
The following sampling steps were carried out independently for 2003 and 2013 (Figure 2, top chart).
First, we acquired suitable, fairly cloud-free and temporally-consistent Landsat scenes between June
and August ensuring fully-developed vegetation. The collected Landsat tiles are equally distributed
across the study area to capture all land cover entities (Figure 4). For each year, we determined 14
Landsat images. The next step involved retrieving spectrally- and texturally-homogeneous objects.
This has been achieved with the Orfeo ToolBox (OTB), a library for processing remote sensing
imagery. Within the OTB, a segmentation application is included, which allows for the multispectral
segmentation of raster images. We included all spectral Landsat bands, as well as the NDVI in the
segmentation process and applied the implemented mean-shift segmentation algorithm, grouping
and generating objects with similar spectral and textural properties. As a minimum sampling unit,
we defined objects covering at least an area of nine MODIS pixel, as suggested by Congalton and
Greene [46]. As a clear visual interpretation of the prevailing land cover is only possible from
high quality and high resolution data, we selected polygon objects embedded in available VHR
Google Earth data. Nonetheless, the marginal areas of each object might still be influenced by mixed
pixels. The negative buffering of one MODIS pixel removes the potential transitional pixels to further
guarantee the homogeneity of the reference objects.
Remote Sens. 2016, 8, 186
8 of 25
Table 2. Name and description of all defined land cover classes in the Yellow River Basin.
ID
Class name
Description
1
Evergreen needle-leaved forests (closed)
2
3
Evergreen needle-leaved shrub and woodland (open)
Deciduous broadleaved forests (closed)
4
5
6
7
8
9
10
11
Deciduous broadleaved shrub and woodland (open)
Grassland
Sparse vegetation
One season cropland
Two season cropland
Natural vegetation/agriculture mosaics
Agriculture/natural vegetation mosaics
Aquaculture
12
13
14
15
16
17
18
Wetlands
Water bodies
Tidal flats
Snow and ice
Deserts (sandy)
Bare areas
Artificial areas
Forest with predominant evergreen needle-leaved tree species, covering at least 65%
and height exceeding 2 m
Evergreen needle-leaved forests covering 15%–40% and mixed with grassland and/or shrub entities.
Forest with predominant broadleaved deciduous tree species, covering at least 65%
and height exceeding 2 m
Deciduous broadleaved forests covering 15%–40% and mixed with grassland and/or shrub entities
Herbaceous vegetation layer with less than 10% woodland and shrub coverage
Sparse shrub and herbaceous vegetation entities covering 5%–15%
Agricultural areas with one harvest per year
Agricultural areas with two harvests per year
Predominant natural vegetation entities (grassland, shrub, woodland), accompanied with cropland
Predominant cropland, accompanied with natural vegetation entities
Water ponds used for aquaculture production, mainly for fish, crustaceans and turtles,
usually surrounded and intersected with grassland
Areas saturated with salt or fresh water with a permanent mosaic of water and herbs or woodland
Areas covered with either fresh or salt water
Coastal wetlands exposed to tidal amplitude consisting of unconsolidated sediments
Areas permanently covered by snow or ice
Barren area covered by sand dunes
Barren land with natural vegetation less than 5%
Built up areas and associated areas
Figure 4. Location and extent of Landsat (red) and very high-resolution (VHR) imagery (blue)
footprints for 2003 (top) and 2013 (bottom). The table below shows meta-information for each
respective Landsat image.
Remote Sens. 2016, 8, 186
9 of 25
To ensure an unbiased distribution of objects across the study area, we randomly generated
2500 points and selected the corresponding polygon. Simple random sampling is the least
biased sampling technique, because all parts of the study area have an equal chance of being
selected. For each land cover class, we decided to select at least 100 samples, as suggested by
several authors [46–48], to get a representative and statistically-valid representation of the complex
landscape. A larger sampling size has been gained for dominant land cover classes and those classes
with inherent variability and complexity [49]. Finally, we assigned the class-specific ID code to
each polygon by visual interpretation. The drawback of simple random sampling is, however, that
small and rarely-occurring areas may be under-represented, because they do not ensure a uniform
distribution of points across the landscape. This problem has also appeared in our sampling strategy.
We therefore manually sampled underrepresented land cover classes until at least 100 samples had
been taken. For all collected samples, we kept a minimum sampling distance of at least 30 MODIS
pixels to avoid spatial autocorrelation effects and removed all objects not fulfilling this criterion.
The final reference dataset for training and testing contained a total of 2340 objects for 2003 and 2232
for 2013. The distribution of each reference dataset can be obtained in Figure 4.
3.4. Classification Approach, Post-Classification and Accuracy Assessment
Various classification models for land cover applications have been used, such as the generalized
linear model (GLM), support vector machine (SVM) or decision tree learning, among others. Decision
tree-based classifiers, such as random forests (RF), have been successfully and extensively applied to
land cover studies and tend to yield higher accuracies than other conventional classifiers, including
maximum likelihood and standard trees [50–52]. Aside from the high performance, RF is robust to
noise and overfitting, and the algorithm is comparably quicker than methods based on boosting [53].
Further, the handling of the classifier is user-friendly, as the model can be run without much tuning
of the parameters. In this study, the entire RF classification was carried out using R and the
package randomForestimplementing Breiman’s random forest algorithm [54]. RF algorithms are used
to construct multiple decision trees, which are built independently, using a bootstrap sample of the
dataset [53]. RF searches for a random sample of the predictors and chooses the best split amongst
the predictors. The final classification is based on the majority vote of the ensemble. Here, we built up
an RF model with 300 individual trees, as prior testing showed a convergence of the mean squared
error (MSE) at that value. For further model tuning and building, we applied the R package caret
developed by Kuhn et al. [55], resulting in a final mtryvalue of 15 for both years. This value describes
how many variables are sampled at each split. The RF classifier ran independently for 2003 and 2013.
Caret integrates a measure of predictor importance to see which variable were the most informative in
making distinctions between the land cover classes. This measure is scaled from 0–100. To generate
an unbiased and stable selection of important and non-important variables prior to fitting the RF
model, the Boruta algorithm was applied [56]. This algorithm tests if variables are expected to be less
important than random samples. No variable was deemed to be non-important.
The thematic accuracy of the two final land cover products was assessed by splitting the
comprehensive reference dataset polygon-wise into a 30% proportion set. The remaining 70% served
as the training samples. We chose the 70/30% split option as it performs best using RF algorithms,
as demonstrated by Adelabu et al. [57]. The use of confusion matrices is a clear way to present the
accuracy, which contains class-specific user’s, producer’s and overall accuracy [46,58].
Many land cover classes that we defined show similar spectral and temporal responses, making
their exact classification ambiguous. Therefore, several logical post-classification rules using ancillary
data were defined to improve the classification accuracy (Figure 5). For instance, mountain shadows
are often misclassified as water. To correct for this issue, the slope calculated from a DEM was
superimposed, and all water bodies with an inclination >1◦ were reclassified to the bare class, as
lentic water is not present on slopes. The spectral differentiation between barren land on steep
slopes and artificial surfaces was revealed to be a challenge. The auxiliary Nighttime Lights Time
Remote Sens. 2016, 8, 186
10 of 25
Series (Version 4) derived from the DMSP Operational Linescan System (OLS) instruments served
as an indication for urban and built-up structures [59]. As suggested by Small et al. [60], we
applied a threshold of >86% to detect lighted urban areas. Double-cropping cycles of wheat-maize
or wheat-soybean are very common in the Yellow River Basin, particularly in the North China Plain,
where favorable environmental conditions are present [61]. As additional output, TIMESAT delivers
a mask with information about seasonal cycles, and we reclassified the agricultural class accordingly.
Figure 5. Post-classification decision rules for the Yellow River Basin land cover classification.
4. Results
The result are presented in three main parts; for the first of which, we highlight the novel
bi-temporal land cover products for the Yellow River Basin for 2003 and 2013 and describe the current
land cover features, focusing specifically on each ecoregion. Additionally, we reveal land dynamics
throughout the basin that have occurred in the last decade. Local hotspot regions of change are
shown, focusing on different land cover types. Next, we compare our derived product to existing
global and national land cover datasets by selecting local case study sites representing some of the
basin’s complex land cover structure.
4.1. Current Land Cover Characteristics and Dynamics in the Yellow River Basin
The classification outcome of the Yellow River Basin land cover products (YRB LC 2003 and 2013)
is presented in Figure 6, with an overall accuracy of 87% and 84% for 2003 and 2013, respectively.
We computed class-specific user’s and producer’s accuracy, which are summarized in the confusion
matrices in Table 3. The generated products depict a detailed extent and distribution of the current
land cover status and dynamics, not presented in previous land cover studies thus far. We adapted
and defined the classes specifically for the Yellow River Basin, including classes of natural vegetation
(terrestrial and aquatic), cultivated classes, mosaic classes, and non-vegetated, artificial classes,
resulting in a total of 18 land cover classes. The land cover proportions and their changes between
2003 and 2013 for the basin’s four major geographical units, the Qinghai-Tibet Plateau, Ordos Plateau,
Loess Plateau and North China Plain, including the delta region (Figure 1, top), can be obtained from
Figure 7. The land cover proportions for the entire basin for 2003 and 2013 are summarized in Table 4.
The land cover forms show significant variation across the basin, as can be seen in Figure 7. The
vast elevated Qinghai-Tibet Plateau is dominated by alpine grasslands covering around 65% in 2013.
Vegetation form and land cover proportions change significantly in the Ordos Plateau steppe, where
sparsely-vegetated and barren areas are predominant. Agricultural production is only possible along
the river banks. The Loess Plateau, natively covered by extensive deciduous forest structures, is
presently a major area for agriculture cultivation. The degree and extent of cultivation processes
throughout the basin climaxes in the fertile and vast North China Plain, where a greater part of the
land surface is dominated by agriculture.
Remote Sens. 2016, 8, 186
11 of 25
Figure 6. Land cover classification products for the Yellow River Basin (YRB LC) for 2003 (top) and
2013 (bottom). Framed areas in the top panel mark the regions with significant land cover changes
shown in Figure 9. The subsets in the bottom panel were used for comparing our Yellow River
classification product with existing global products depicted explicitly in Figure 10.
Figure 7. Land cover proportions for each ecoregion and year (2003, left; 2013, right). The color code
of the different land cover types conforms to the colors introduced in Figure 6. The coverage of each
ecoregion can be seen in Figure 1.
Remote Sens. 2016, 8, 186
12 of 25
Classification
Classification
Table 3. Confusion matrix showing user’s and producer’s accuracy for each class and overall accuracy
in percent for 2003 (top) and 2013 (bottom).
2013
1
2
3
4
5
6
7
9
10
11
12
13
14
15
16
17
18
1
89.15
14.18
3.61
0.22
0
0
0
0
0
0
0
0
0
0
0
0
0
2
9.52
82.53
1.45
3.23
0.60
0.72
0
0
0
2.5
0
0
0
0
0
0
0
3
0
0.46
92.29
0
0
0
0
0
0
0
0.43
0
0
0
0
0
0
4
1.06
2.30
1.69
78.06
7.01
0
0
7.6
1.13
0
0
0
0
0
0
0
0
5
0.26
0.23
1.3
3.36
69.14
1.67
0.45
7.04
0.91
0
0
0.23
0
0
0
0.23
1.21
6
0
0
0
0
0
83.73
0
0.29
0.45
0
0
0
0
0
1.12
5.25
5.84
7
0
0
0
1.29
0
0.48
82.21
4.40
8.14
0
1.10
0
0
0
0
0
0
9
0
0
0
5.26
7.21
0
5.63
79.77
1.58
0
0
0
0
0
0
0
0
PA (%)
81.01
84.47
90.97
80.67
84.14
85.99
85.28
76.40
2013
1
2
3
4
5
6
7
9
10
11
12
13
14
15
16
17
18
1
94.04
5.91
0.48
1.66
0
0
0
0
0
0
0
0
0
0
0
0
0
2
4.70
89.83
1.45
13.39
0.92
0
0
0.44
0
0
0
0
0
0
0
0
0
3
1.25
0.47
91.04
6.98
0.23
0
0
0.22
0
0
0.53
0
0
0
0
0
0
4
0
0.95
5.57
65.96
10.30
0
0.25
5.01
0.84
0
0.26
0
0
0
0
0
0
5
0
2.36
0.48
2.99
64.32
0.97
1.27
8.71
2.51
0
0
0
0
0.25
0
0
0
6
0
0
0
0
0.23
71.07
0
0.65
0.42
0
0
0
0
0
1.26
8.88
3.05
7
0
0.24
0
0
5.72
1.94
90.93
1.09
9.41
0.48
2.65
0
0
0.23
0
0.51
.910
9
0
0.24
0
11.87
5.95
0.39
1.01
76.25
12.55
0
0.26
0
0
0
0
0
0
PA (%)
90.36
84.07
92.38
71.23
78.68
86.73
77.48
71.57
Reference
10
11
0
0
0
0
0
0
3.2
0
5.01
0
0
0
10.14 0.23
0.3
0
87.10
0
0
94.00
0
0.82
0
11.9
0
0
0
0
0
0
0
0
0
0.25
12
0
0
0
0
5.21
0
1.13
0
0
0
97.3
0.35
3.5
0
0
0
0
13
0
0
0
0
0.20
0
0
0
0
5.99
0.55
82.00
0.23
1.05
0
0
0.73
14
0
0
0
0
5.01
0
0
0
0
0
0
0
96.26
0
0
2.94
3.65
15
0
0
0
0
0
0
0
0
0
0
0
5.69
0
98.94
0
3.15
0
16
0
0
0
0
0
0
0
0
0
0
0
0
5.69
0
98.92
0
3.15
17
0
0
0
0
0.60
3.59
0.23
0
0
0
0
0
0
0
13.09
77.73
0.49
18
0
0
0
0
0
6.46
0
0.29
0
0
0
0
0
0
0
5.46
87.59
UA (%)
89.15
82.53
92.29
78.06
69.14
83.73
82.21
79.77
87.10
94.00
97.30
82.00
96.26
98.94
98.92
77.73
87.59
87.31
88.33
90.68
88.41
92.16
90.23
82.96
86.96
86.83
Reference
10
11
0
0
0
0
0.24
0
1.06
0
9.34
4.15
0.19
0
1.01
3.54
6.97
0
74.27
0
0
92.36
0
2.38
0
12.27
0
0.27
0
1
0
0
0
0
0
0.30
12
0
0
0.73
0
0
4.47
0.76
0
0
0
92.59
0
0
0
0
0.91
0.80
13
0
0
0
0
0.23
0.19
0.51
0
0
5.97
0.26
87.50
0
0
0
1.27
2.13
14
0
0
0
0
0
0
0
0
0
1.19
0.53
0
98.94
0
0
2.54
2.13
15
0
0
0
0
0
0
0.19
0
0
0
0
0
0
98.15
0
2.1
0
16
0
0
0
0
0
1.75
0
0
0
0
0
0
0
0
97.47
2.5
0
17
0
0
0
0
0.69
12.62
0
0.22
0
0
0
0
0
0
1.27
76.14
1.22
18
0
0
0
0
0.12
6.21
1.10
0.44
0
0
0.53
0
0
0
0
9.4
91.50
UA (%)
94.04
89.83
91.04
65.97
64.32
71.07
90.93
76.25
74.27
92.36
92.59
87.50
98.94
98.15
97.47
76.14
91.50
90.91
90.17
90.17
98.30
95.54
79.37
81.08
84.31
81.57
80.32
79.30
Table 4. Total area in km2 and in the percentage of the total area for each land cover class and year.
2003
Class ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
2013
Area (km2 )
Area (%)
Area (km2 )
Area (%)
32,663
30,425
261,110
127,876
1,049,087
759,812
895,934
63,480
90,534
153,710
5123
8992
23,797
2003
1319
100,075
44,907
63,676
0.88
0.82
7.03
3.44
28.26
20.47
23.93
1.69
2.44
4.14
0.14
0.24
0.64
0.05
0.04
2.70
1.21
1.72
33,992
25,828
271.269
135,607
1,053,858
723.970
902,009
15,582
199,671
46,135
7269
7752
22,427
2782
1359
92,849
44,004
87,332
0.92
0.70
7.31
3.65
28.39
19.50
25.37
0.42
5.38
1.24
0,20
0.21
0.60
0.07
0.04
2.50
1.19
2.35
4.1.1. The Qinghai-Tibet Plateau
The Qinghai-Tibet Plateau is the world’s highest and largest grassland plateau and feeds many
of Asia’s most important rivers. With altitudes exceeding 4500 m, the plateau is characterized by
harsh environmental conditions with low temperatures and little precipitation. The landscape is
dominated by relatively homogeneous alpine grassland tundra accompanied with areas of spare
vegetation and bare areas. The homogeneous grassland systems achieved in this region higher user’s
(89%) and producer’s (91%) accuracies compared to, e.g., the Loess Plateau, where accuracies are
Remote Sens. 2016, 8, 186
13 of 25
considerably lower. Around 65% of the plateau is covered with grassland, as can be seen in Figure 7.
This figure has not changed in the last 10 years. A marginal increase of sparse vegetation cover
at the costs of grassland is detected in the vicinity of the two freshwater lakes, the Gyaring and
Nyoring lakes. At the eastern foothills and in river valleys with milder climatic conditions, open and
closed needle-leaved forest patches can be found. Glacier remnants are present in the ranges of the
A’nyê Magên Mountains. The harsh environmental conditions make crop cultivation rather difficult.
Only in low altitude valleys with milder climates are farmers able to grow wheat and vegetables in
small-scale and subsistence agriculture.
4.1.2. The Ordos Plateau Steppe
Northward, there lies the extensive desert and steppe-like Ordos Plateau. Because of the arid
conditions, dry-adapted and resistant xeric shrubs and grassland vegetation become more dominant,
covering the surface very sparsely. The lack of precipitation in conjunction with non-nutritious soils
inhibit productive cultivation of land for agriculture. Exceptions are small scattered oases throughout
the steppe, along the banks of the Yellow River and large-scale irrigation districts, where sufficient
water is available to grow crops. Irrigation farming is practiced in two major irrigation districts: the
Qingtongxia district in the west and the Hetao district to the north. The main cultivated crops in this
region are corn and wheat [62]. The Ordos region also comprises two large sandy deserts covering
an approximate area of 6%: the Kubugi Desert in the north and the Maowusu Desert further south,
with sparse vegetation along the sandy dunes. Since 2003, the proportions of sparse vegetation have
declined by approximately 17%, being replaced mainly for agricultural purposes and for areas of
mixed grassland/cropland mosaics. Further, urban and built-up areas have significantly expanded,
particularly along the river banks.
4.1.3. The Loess Plateau
In contrast to the Ordos Plateau, the Loess Plateau is attributed with fertile, but highly erosive
soils, allowing for cultivation. Once covered by dense vegetation, dominated by temperate open
and closed deciduous (mainly oak, birch) and in higher altitudes by evergreen forests (spruce,
mountain-ash), a long history of deforestation and over-grazing has turned the Loess Plateau into
a degraded region with high erosion rates [63]. Many Yellow River tributaries originate on the
central Loess plateau and cut the hilly area into numerous narrow ridges and mounds, making
the surface rugged and uneven. At the eastern margins of the basin, the Loess Plateau consists
of still extensive mixed deciduous forests in the Lu Liang Mountains, which gradually changes
south-eastward to low steppe vegetation. In 2013, forest and woodland structures (open and closed)
share an area of around 15,000 km2 , equal to 30% of the Loess Plateau, and this has not changed
significantly during the last decade. The central part is characterized by small-scale mosaics of
cropland and natural vegetation and areas of sparse vegetation, particularly on sloping areas. Until
2013, the proportion of natural vegetation and mixed cropland patches has significantly increased,
while sparsely-vegetated areas have dwindled. Particularly, the mosaic class, having agriculture as
the major land form, was transformed to the class “natural/agriculture vegetation mosaics”, with a
predominant share of natural vegetation. Many land policies and conservation programs have been
launched, implementing terracing and greening the slopes to mitigate and stop the massive silt loads
entering the Yellow River. The booming economy, mainly based on coal and chemical production,
has led to an increase in urban and built-up areas in this region. Today, 50 million people populate
the entire Loess Plateau.
4.1.4. The North China Plain and Delta
Further south, the North China Plain commences with the wide and densely-populated valley
of the Wei River, the largest tributary of the Yellow River, which embodies one of the oldest irrigation
districts in China. Through canals, the water is distributed across the valley, supplying major cities,
Remote Sens. 2016, 8, 186
14 of 25
which considerably sprawled during the last decade by 24%; a similar figure was suggested by [1]
(Figures 6 and 7 and Table 4). The Qinling Mountain Range delimits the Yellow River Basin to
the south. The dominant vegetation forms are temperate deciduous forest ecosystems. Following
the Yellow River eastwards towards the Bohai Sea, the vast alluvial region is dominated by a high
level of intensified agricultural production, producing wheat, corn, sorghum and cotton, covering
around 66%. Between 2003 and 2013, the cultivated area increased by 5%, which is by far not the
order of magnitude of the crop output, which increased by 22% [1]. As suggested by Li et al. [61],
cropping intensity is achieved by switching from single- to double-cropping cycles. This is obviously
not the case in the Yellow River Basin, where cropping cycles decreased considerably, which is in
accordance with data provided by the National Bureau of Statistics of China, indicating a reduction
of two seasonal crops. Henan and Shandong are the most populous provinces located in the
Yellow River Basin.
The Yellow River Delta is considered one of China’s key regions in terms of economic
development and has intensely evolved during the last few decades. The fertile soils make intensified
agricultural crop cultivation possible. Furthermore, the high abundance of oil and gas resources
triggered a rapid industrialization and urbanization in this region. Along the coastline, many ponds
emerged for aquaculture practices and are still expanding (increase by 14%). All of this has harnessed
the ecologically-sensitive delta area, one of only a few remnants of wetland ecosystems that remain,
which are vitally important to local and even global avian biodiversity. To protect this sensitive
area, the Yellow River Delta Nature Reserve was founded and consistently expanded during the last
few years.
4.2. Variable Importance
The importance measures for 2003 and 2013 of all 37 MODIS predictors for land cover distinction
in the Yellow River Basin are depicted in Figure 8, ordered from most to least important. Both years
show similar patterns. The seasonal and small percentiles (10% and 25%) for both spectral bands and
NDVI rank amongst the most relevant metrics in delineating the land cover for the basin. Base level,
amplitude, peak value and the NDVI integral are found to be the most important seasonal metrics,
while left derivative, middle season and the large integral remain least important. Likewise, higher
percentiles (50% and 75%), as well as slope and altitude play a minor role in land cover delineation.
Figure 8. Importance of each predictor variable used in the random forest classifiers for 2003
(left panel) and 2013 (right panel), scaled from 0–100.
Remote Sens. 2016, 8, 186
15 of 25
4.3. Local Dynamic Hotspot Regions in the Basin
The Yellow River Basin has undergone tremendous land cover dynamics during the last
decade, triggered by recent socio-economic developments, such as the growing economy, increasing
urbanization, agricultural expansion and intensification, and various land use policies that have been
launched in the past decade. In this section, we highlight the most significant dynamics exemplarily
by means of subsets, as depicted in Figure 9.
Figure 9. Local hotspots of change within the Yellow River Basin. The numbers refer to the exact
geographical location as indicated in Figure 6. The color code of the different land cover types
conforms to the colors introduced in Figure 6.
Remote Sens. 2016, 8, 186
16 of 25
4.3.1. Urban Expansion
Despite population growth rates in China remaining more or less constant since 2003, the
urbanization rate has almost doubled in this time [1]. This is also true for the Yellow River provinces,
where urban areas of the major cities have expanded with an annual rate of up to 4%. Xi’an is the
capital of Shaanxi province, which is one of the oldest cities in China, comprising a total present-day
population of around 8.5 million, and growing annually by more than 3% (Figure 9 1). The city is
embedded in the floodplains of the Wei River, a major region for agricultural production. As part of
China’s so-called “National Economic and Technological Development Zones”, Xi’an is a center for
high-tech industry and is specifically tailored to attract foreign investments and funds to foster local
economies and employment. Artificial structures have further expanded mainly along river banks
and in the major irrigation districts, where intense agricultural production is present. Throughout
the basin, urban and built-up areas increased by almost 35% during the last decade.
4.3.2. Agricultural Encroachment
Agriculture is the dominant landscape feature in the Yellow River Basin, covering one third
of the entire basin area. Between 2002 and 2013, the absolute numbers have not changed much,
and there is even a noticeable decline in agricultural areas by 5%, which is in concordance with
the national statistics [1] and can be seen in Table 4. Despite this overall decline, new areas for
agricultural production have been developed locally in various regions. One example is located in
the arid northern part of Ningxia province (Figure 9 2), where the establishment of new irrigation
facilities has improved and intensified agricultural yields. Other examples of locally-encroaching
agricultural fields are areas on the Ordos and Loess Plateau in Shaanxi province and Inner Mongolia,
where the emergence of small-scale farming practices in groundwater-fed oases is presently located in
the Ordos Desert. Another region experiencing significant agricultural dynamics and intensification
is the Yellow River Delta region, mainly at the cost of the prevailing wetlands.
4.3.3. Ecological Restoration
Many decades and even centuries of intense human influence have turned particularly the
Loess Plateau into a degraded landscape, where natural forests and shrublands in most areas were
largely replaced centuries ago by agricultural cropland and pasture, contributing to recent severe soil
erosion. In order to protect soils and maintain the fertility of such areas, diverse land use policies and
conservation management plans have been launched to impede desertification, increase ecological
conditions and sustain agricultural production, but in a more sustainable way. In the late 1990s,
the Chinese government launched a nationwide scheme called the “Grain-for-Green” program, with
the goal to increase and restore natural vegetation coverage on steep slopes of the Loess Plateau [64].
The change of the landscape structure as a result of this can be nicely seen in Figure 9, where grassland
areas expanded on previous cropland or sparse land. The selected area is embedded in the Shanbei
region (Shaanxi province).
In 1996, by the World Bank initiated the “Loess Plateau Watershed Rehabilitation Project” as one
of the largest restoration programs worldwide and targeted specifically degraded areas with a low
degree of vegetation coverage and susceptible to soil erosion. The primary objectives of this project
were to create sustainable crop production on fertile arable land on terraced gulch slopes and to
further stabilize land on slopes by planting a range of trees, shrubs and grasses to prevent soil
degradation and erosion and the loss of valuable land to desertification. One target area is depicted
in Figure 9, located in Gansu province, showing the successful recovery of grassland systems.
4.4. The YRB LC Product vs. Global and National Land Cover Products
Global land cover products are designed to delineate global land cover patterns, but their
suitability for regional applications is not always given. We compare our regional land cover product
Remote Sens. 2016, 8, 186
17 of 25
LC YRB 2013 to the MODIS MCD12Q1 2012 [19] (IGBP), the GlobCover 2009 [18] and the ESA CCI-LC
2010 [20]. Further, a high resolution Landsat 8 OLI scene (30 m) from 2013 was taken as reference
imagery, showing finer prevailing landscape features in higher spatial detail. For this comparison,
we have chosen three subregions, which characterize and represent the major landscape features in
the Yellow River Basin determined by rather complex and heterogeneous conditions (Figure 10). The
first subset (Row 1 in Figure 10) focuses on a semi-natural and agriculturally-dominated landscape
with disjunct forest patches. The Yellow River Delta is the second subset, also a human-dominated
landscape with biodiverse wetlands at the river mouth (second row). Further, we selected a rather
natural ecosystem, the vast alpine grassland structures on the Qinghai-Tibet plateau presented in
Column 3 of Figure 10. The geographical extent and distribution of each subregion is depicted
in Figure 6.
The first region (Row 1 in Figure 10), located in the central part of Shaanxi province in the
floodplains of the Wei River, covers disjunctive forest patches (deciduous and evergreen) surrounded
by intensified agricultural areas. Transitional and fragmented areas of semi-natural vegetation are
present at the forest edges, where small-scale farming practices cut the forest into small mosaics.
Urban and peri-urban structures are scattered across the Wei floodplain, close to the capital of Shaanxi
province (Xi’an). This complex landscape pattern is reflected in high spatial detail by the YRB LC
2013, which includes even the small-scale mosaic structure of natural and semi-natural vegetation
patches. The continuous forest areas are delineated in all considered land cover maps. The MCD12Q1
tends to overestimate deciduous forest formations, and it fails to differentiate between forest types.
Despite the lower class accuracies for open shrub- and wood-land and for the mosaic classes,
their distribution is clearly discernible in the YRB LC 2013 map, considering this heterogeneous
semi-natural landscape. Only ESA CCI-LC reflects the tessellated pattern in similar detail. Looking
at the artificial class, our YRB LC 2013 map is exclusively able to delineate complex and small-scale
urban structures.
In terms of complexity, a similar landscape is the Yellow River Delta, which is a region that
has been intensively shaped by human interactions (Row 2 in Figure 10). The delta consists of
very dynamic, temporally-flooded tidal flats and wetlands at the river mouth. Large aquaculture
ponds intersect the near coastal zone and produce mainly fish and crustaceans [7]. The inner delta
mainly consists of agricultural fields and many small scattered urban areas. Dongjing is the major city
located in the southern delta region. Aquaculture ponds are depicted in the high resolution Landsat
imagery and are congruent with the classified areas in YRB LC 2013. The other considered land cover
maps clearly identified aquaculture incorrectly as either grassland, rainfed cropland or water bodies.
While wetlands are often misclassified as grassland or irrigated cropland, cropland areas are depicted
correctly in almost all land cover maps, except for ESA CCI-LC, where evident cropland areas are
assigned to grassland. It can be further noted that urban areas seem to be better detected compared
to the previous subset. GlobCover still greatly underestimates urban structures. Once again, YRB LC
2013 delineates these dispersed areas in higher spatial detail.
The third subregion, shown in Row 3 of Figure 10, covers rather homogeneous alpine grassland
steppes (Qinghai province) interspersed with barren and sparse mountain ranges, showing low
vegetation coverage. Two large freshwater lakes and small brackish water bodies are scattered across
the area. Despite this homogeneity, the considered maps show different classification schemes, except
for water bodies, which are delineated well in all maps. From the Landsat imagery, it is evident
that YRB LC 2013 follows the landscape pattern of the high resolution image. The MODIS product
classifies all terrestrial areas consistently as grassland and neglects sparse vegetation structures
on mountainous areas. GlobCover captures barren structures, but assigns grassland steppes to
mosaic croplands/vegetation and irrigated cropland classes. Furthermore, ESA CCI-LC introduces
cropland-related classes in this area, even though agricultural production is not realistic in this harsh
environment (altitude 4300 m) with its infertile soil and low temperatures.
Remote Sens. 2016, 8, 186
18 of 25
Figure 10. Comparison of Landsat 8 imagery (RGB band combination: 7-5-3; Column 1), against the YRB LC 2013 product (Column 2), MODIS MCD12Q1 2012
(Column 3), GlobCover 2009 (Column 4) and the ESA CCI-LC (Column 5) products. We selected three representative and rather heterogenetic subregions within
the Yellow River Basin: a tessellated area with forest patches intersected with cropland in the south (Row 1); the delta area accompanied by urban, agriculture and
wetlands (Row 2); and the source region dominated by grassland and sparse vegetation mosaics (Row 3). Color code for YRB LC 2013 as introduced in Figure 6.
Remote Sens. 2016, 8, 186
19 of 25
5. Discussion
Prior work has highlighted the necessity of detailed, consistent and multi-temporal regional
landscape information, making it important for climate or hydrological modeling, as well as for
various decision-making processes. Hence, we developed in this paper novel land cover products for
the very dynamic Yellow River Basin by revealing the current landscape characteristics and assessing
past changes throughout the basin. A straightforward classification procedure was developed
using high temporal moderate spatial resolution MODIS data, which enable accurate delineation of
heterogeneous landscapes at large spatial scales.
5.1. Land Cover Classification and Dynamics
The two maps presented in this study describe the land cover status for the Yellow River Basin
for the years 2003 and 2013. They provide thematic detailed and regionally-adapted land cover
classes derived from high-temporal MODIS data at 250-m spatial resolution. For classification,
we took spectral, phenological, temporal and ancillary information into consideration. In line
with many studies using phenological information for land cover differentiation [24–26,29,65,66],
we demonstrated the advantageous application of such data, particularly for regions comprising
complex and heterogeneous environmental conditions. The Yellow River watershed covers an
approximate area of 750,000 km2 at an altitudinal gradient of more than 4600 m, therefore creating
very diverse bioclimatic and environmental conditions.
The most obvious dynamics includes the landscape alteration of the Loess and Ordos Plateau,
where land use policies and local conservation management plans combat the environmental
deterioration by restoring vegetation cover and various ecosystem services, such as carbon
sequestration, hydrological services and soil retention. As already shown in Sections 4.1 and 4.2,
our results revealed changing ecological conditions, with the revegetation of barren and sparse land
and transformation of sloped cropland to grass or woodland. Our results are analogous to a number
of local remote sensing and landscape analysis studies, confirming an ecological and environmental
enhancement also at larger scales [4,11,67–69]. Studies investigating the net primary productivity
(NPP), a crucial proxy for vegetation productivity and carbon fluxes, indicated a steady increase over
the last decade [70,71] stemming from land conservation programs. The study of Zhai et al., 2015,
emphasizes the strong anthropogenic effect on this landscape’s changes and excluded precipitation
as the driving factor, showing no significant correlation between precipitation and greening. Beside
the emphasis on nature restoration, social facets also fall into the scope of the respective programs,
aiming towards a reduction of the impoverishment of the local people.
Agriculture is a major landscape facet in the Yellow River Basin, and cultivation practices
started centuries, even millennia ago. The Chinese government promotes the objective of food
self-sufficiency in order to reduce the reliance on international trade markets [72]. Given the growing
demand for food by a rising and increasingly affluent population, the agricultural sector is facing
serious challenges to maintain food security, particularly considering progressive degradation of
land and water resources. Absolute numbers of cultivated areas within the basin remained stable,
while the total annual yield of agricultural commodities has tremendously increased [1] by means
of improved farming practices, irrigation, fertilizer and pesticide consumption and more advanced
machinery [73–76]. For instance, chemical fertilizer consumption has risen substantially by 10% in
Shandon and even 32% in Henan province, the major production zones in the basin [1]. However,
agricultural areas are still dynamic features across the basin, and we see certain localized changes.
Declines in cultivated areas occurred mainly in Shaanxi, where the “Grain for Green” program took
large areas out of cultivation [77]. With expanding and more efficient irrigation facilities, crop areas
developed more strongly in arid regions, such as Northern Gansu or Ningxia, where the Yellow
River and its tributaries provide a sufficient amount of water for irrigation. Multi-cropping harvest
cycles per year are widespread in China to increase cropping intensity and yields without expanding
the cultivated area. Located in the floodplains of the Wei River and North China Plain, these
Remote Sens. 2016, 8, 186
20 of 25
multi-crop cycles mainly produce corn, wheat, oil-bearing crops, vegetables and fruits. Our results
indicate a considerable decline of two seasonal cropping practices. Agricultural statistics from the
National Bureau of Statistics reveal a decrease in total sown areas of winter wheat, replaced by more
profitable crops, such as fruits or oil crops, which are showing an inverse trend [1]. However, the
tremendous drop of two-season crops is not fully reflected in the statistics. There are various crop
systems practiced in this region, periodically rotating from three-crop harvests in two years to annual
harvests [78].
Urban and built-up areas have sprawled at tremendous speed, encroaching on surrounding
lands covered by agriculture or natural vegetation. The economically most strongly developed
provinces Shandong and Henan encompass the highest urban population exceeding 60% at annual
growth rates up to 4%. Similar trends show less developed provinces, e.g., Qinghai or Ningxia,
even at higher relative urbanization rates (10%). More and more peasants are moving to cities as
those are the loci of employment and opportunities. Besides the urban sprawl of cities, built-up
and artificial areas reveal significant encroachment along the river banks, where local industries and
infrastructure emerged. Growing urbanization accompanied with a more affluent population usually
leads to higher energy demands and consumption, mainly built on coal, which currently supplies
two-thirds of China’s overall energy use [1]. Coal production for domestic and international markets
across the Yellow River Basin has increased by 25% in Shanxi province between 2003 and 2013, the
major coal production center in the basin, and even by 300%, albeit from a low base [1]. This is
reflected in our classified maps, where we can see localized emergence and expansion of mining
areas and an increase of barren and impervious built-up land on the Ordos Plateau. Opencast mining
activities have recently emerged and expanded in Qinghai province, replacing fragile and sensitive
alpine grassland meadows.
5.2. Land Cover Accuracies
Given the very complex and heterogeneous landscape characteristics and the vast geographical
extent, the attained thematic accuracies of 87% (2013) and 84% are satisfactory and comparable
to other equally-complex regional land cover studies across the globe [24,26,29,45,65]. Seasonal
information, based on high-temporal time-series data, is deemed to be a very important input feature
in delineating different land cover classes, which can be seen in Figure 8. As illustrated in Table 3,
the class-specific accuracies are in a high range, but classification challenges remain. Both mosaic
classes and open deciduous broad-leaved shrub and woodland show the highest confusion, mainly
with grassland, possessing user’s accuracies of 65%–85% and producer’s accuracies of 70%–86%.
The error pattern is analogous for both years. According to our definition of the mosaic and open
structure classes, the respective pixels may contain a significant area of grassland, making them quite
similar in terms of spectral-temporal properties and, thus, leading to a false class allocation. Using
medium-resolution remote sensing sensors, such as MODIS, and the accompanying lower accuracies,
it was therefore necessary to introduce land cover mosaics to improve the thematic representation
of the spatial heterogeneity inherent in the basin. Indeed, the major cause of misclassification is
attributed to mixed pixels with the use of a conventional (hard) pixel classifier, particularly when
the classification is based on medium to coarse resolution remote sensing data [58]. Similarity,
open and closed forest systems, both evergreen and deciduous, often transit into each other and
contribute to confusion between these classes. A fractional cover approach, where land cover
proportions on a sub-pixel level are delineated, could lead to substantial enhancement of classification
performance [25,79,80] for transitional and small-scale landscape structures, such as those occurring
on the Loess Plateau.
Contrary to this, closed deciduous forests that occur in large-scale continuous patches have
distinct spectral and temporal properties and were mapped with high accuracies exceeding 90%.
Generally, spatially-homogeneous land cover types tend to have higher chances of being classified
successfully. Analogically, the classes snow and ice, water bodies, tidal flats, wetlands and desert
Remote Sens. 2016, 8, 186
21 of 25
yielded high class-specific accuracies with user’s and producer’s accuracies >90%, which usually
encompass very distinct spectral and temporal signals and can be well mapped. Misclassification,
albeit to a lower degree, appears in delineating aquacultural ponds with water bodies, barren land
with deserts and barren areas with artificial and built-up surfaces. The latter case is an ambitious
classification task, requiring the application of ancillary information. In this respect, we applied
the Nighttime Lights Time Series for 2003 and 2013, which substantially helped to improve the
final results.
6. Conclusions
This study presented a novel, consistent, thematically highly detailed and bi-temporal land cover
analysis representing the heterogeneous land cover characteristics for the Yellow River Basin in China
between 2003 and 2013 and revealing major changes in this dynamic region. Working at a spatial
resolution of 250 m, we used the advances of phenology-based metrics derived from high-temporal
MODIS time series to discern different land cover classes by their temporal trajectory at large spatial
scales. Despite complex and challenging land surface characteristics, which include a wide range
of land cover entities with different environmental conditions, the achieved thematic accuracies of
87% (2003) and 84% (2013) are promising. Although the tessellated land cover classes showed the
lowest accuracies, the incorporation is necessary to give a better and more realistic representation of
the spatial heterogeneity in the basin.
In conclusion, the presented classification scheme we presented, using spectral, seasonal and
ancillary information, is a straightforward way to provide an enhanced picture of the prevailing
landscape characteristics and dynamics on large spatial scales. To highlight the superiority of the
regional-adapted maps, we compared the YRB LC product to the often used global land cover
maps and clearly demonstrated that these optimized maps are better suited for depicting regional
landscapes. Additionally, multi-temporal information allows for the spatial depiction of trends and
recent land change across the Yellow River Basin. The major identified dynamics include agricultural
and industrial intensification, as well as growing urban and peri-urban areas. Major conservation
policies have considerably shaped the landscape characteristics and improved environmental and
ecological conditions in the last decade. The developed highly detailed land cover maps provide
essential information to support decision making for sustainable land and water management plans
and are of particular interest for national, provincial and local stakeholders. Further, this information
can be used as primary inputs for hydrological or climate modeling.
Acknowledgments: This study was undertaken in the context of the DELIGHT (Delta Information System for
Geoenvironmental and Human Habitat Transition) project, funded by the German Federal Ministry of Education
and Research, BMBF (Federal Ministry of Education and Research), and by the International Science and
Technology Cooperation Program of China, 2012DFG22050. Furthermore, we would like to thank the MODIS
and Landsat science team for providing the data free of charge. Finally, we would like to express our gratitude
to Ursula Gessner for fruitful discussions and helpful insights that greatly helped to improve our manuscript.
Three anonymous reviewers are thanked for greatly improving the quality of this contribution.
Author Contributions: Christian Wohlfart undertook the data precessing and analysis. Claudia Kuenzer and
Gaohuan Liu designed the research scheme and guided the processing and interpretation. Christian Wohlfart
authored the first version of the manuscript, and all authors jointly improved the version based on critical
discussion. Chong Huang supported field work conducted in the Yellow River Basin.
Conflicts of Interest:The authors declare no conflict of interest.
References
1.
2.
3.
NBS (National Bureau of Statistics) China Statistical Database; National Bureau of Statistics: Beijing,
China, 2013.
Ringler, C.; Cai, X.; Wang, J.; Ahmed, A.; Xue, Y.; Xu, Z.; Yang, E.; Zhao, J.; Zhu, T.; Cheng, L.; et al.
Yellow River basin: Living with scarcity. Water Int. 2010, 35, 681–701.
Nakayama, T. Simulation of the effect of irrigation on the hydrologic cycle in the highly cultivated
Yellow River Basin. Agric. For. Meteorol. 2011, 151, 314–327.
Remote Sens. 2016, 8, 186
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
22 of 25
Miao, C.Y.; Yang, L.; Chen, X.H.; Gao, Y. The vegetation cover dynamics (1982–2006) in different erosion
regions of the Yellow River Basin, China. Land Degrad. Dev. 2012, 23, 62–71.
Cai, X. Water stress, water transfer and social equity in Northern China—Implications for policy reforms.
J. Environ. Manag. 2008, 87, 14–25.
Bi, X.; Wang, B.; Lu, Q. Fragmentation effects of oil wells and roads on the Yellow River Delta,
North China. Ocean Coast. Manag. 2011, 54, 256–264.
Ottinger, M.; Kuenzer, C.; Liu, G.; Wang, S.; Dech, S. Monitoring land cover dynamics in the Yellow River
Delta from 1995 to 2010 based on Landsat 5 TM. Appl. Geogr. 2013, 44, 53–68.
Kuenzer, C.; Ottinger, M.; Liu, G.; Sun, B.; Baumhauer, R.; Dech, S. Earth observation-based coastal
zone monitoring of the Yellow River Delta: Dynamics in China’s second largest oil producing region over
four decades. Appl. Geogr. 2014, 55, 92–107.
Dong, L.; Wang, W.; Ma, M.; Kong, J.; Veroustraete, F. The change of land cover and land use and its
impact factors in upriver key regions of the Yellow River. Int. J. Remote Sens. 2009, 30, 1251–1265.
Song, X.; Yang, G.; Yan, C.; Duan, H.; Liu, G.; Zhu, Y. Driving forces behind land use and cover change
in the Qinghai-Tibetan Plateau: a case study of the source region of the Yellow River, Qinghai Province,
China. Environ. Earth Sci. 2009, 59, 793–801.
Chen, Y.F.; Liu, Y.S.; Wang, J.; Yan, J.P.; Guo, X.D. Land use changes of an aeolian-loessial soil area in
Northwest China: implications for ecological restoration. Pedosphere 2009, 19, 356–361.
Guo, W.Q.; Yang, T.B.; Dai, J.G.; Shi, L.; Lu, Z.Y. Vegetation cover changes and their relationship to
climate variation in the source region of the Yellow River, China, 1990–2000. Int. J. Remote Sens. 2008,
29, 2085–2103.
Liang, S.; Ge, S.; Wan, L.; Xu, D. Characteristics and causes of vegetation variation in the source regions
of the Yellow River, China. Int. J. Remote Sens. 2012, 33, 1529–1542.
Zhang, X.; Sun, R.; Zhang, B.; Tong, Q. Land cover classification of the North China Plain using
MODIS-EVI time series. ISPRS J. Photogram. Remote Sens. 2008, 63, 476–484.
Daofeng, L.; Chunhui, L.; Fanghua, H.; Zheng, L. Complex vegetation cover classification study of the
Yellow River Basin based on NDVI data. In Proceedings of the 2004 IEEE International Geoscience and
Remote Sensing Symposium, IGARSS ’04, Anchorage, AK, USA, 20–24 September 2004; pp. 3360–3363.
Wang, S.; Ding, C.; Liu, J. Landscape evolution in the Yellow River Basin using satellite remote sensing
and GIS during the past decade. Int. J. Remote Sens. 2009, 30, 5573–5591.
Bartholomé, E.; Belward, A.S. GLC2000: A new approach to global land cover mapping from Earth
observation data. Int. J. Remote Sens. 2005, 26, 1959–1977.
Bicheron, P.; Defourny, P.; Brockmann, C.; Schouten, L.; Vancutsem, C.; Huc, M.; Bontemps, S.; Leroy, M.;
Achard, F.; Herold, M.; Ranera, F.; Arino, O. GLOBCOVER - Products Description and Validation Report;
MEDIAS-France: Toulouse, France, 2010.
Friedl, M.A.; Sulla-Menashe, D.; Tan, B.; Schneider, A.; Ramankutty, N.; Sibley, A.; Huang, X.
MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets.
Remote Sens. Environ. 2010, 114, 168–182.
Defourney, P.; Kirches, G.; Brockmann, C.; Boettcher, M.; Peters, M.; Bontemps, S.; Lamarche, C.;
Schlerf, M.; Santoro, M. Land Cover CCI Product User Guide Version 2; UCL-Geomatics: Louvain, Belgium,
2014.
McCallum, I.; Obersteiner, M.; Nilsson, S.; Shvidenko, A. A spatial comparison of four satellite derived
1 km global land cover datasets. Int. J. Appl. Earth Obs. Geoinform. 2006, 8, 246–255.
Herold, M.; Mayaux, P.; Woodcock, C.; Baccini, A.; Schmullius, C. Some challenges in global land cover
mapping: An assessment of agreement and accuracy in existing 1 km datasets. Remote Sens. Environ.
2008, 112, 2538–2556.
Alcantara, C.; Kuemmerle, T.; Prishchepov, A.V.; Radeloff, V.C. Mapping abandoned agriculture with
multi-temporal {MODIS} satellite data. Remote Sens. Environ. 2012, 124, 334–347.
Klein, I.; Gessner, U.; Kuenzer, C. Regional land cover mapping and change detection in Central Asia
using MODIS time-series. Appl. Geogr. 2012, 35, 219–234.
Wohlfart, C.; Wegmann, M.; Leimgruber, P. Mapping threatened dry deciduous dipterocarp forest in
South-east Asia for conservation management. Trop. Conserv. Sci. 2014, 7, 597–613.
Remote Sens. 2016, 8, 186
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
23 of 25
Gessner, U.; Machwitz, M.; Esch, T.; Tillack, A.; Naeimi, V.; Kuenzer, C.; Dech, S. Multi-sensor mapping
of West African land cover using MODIS, ASAR and TanDEM-X/TerraSAR-X data. Remote Sens. Environ.
2015, 164, 282–297.
Clark, M.L.; Aide, T.M.; Riner, G. Land change for all municipalities in Latin America and the Caribbean
assessed from 250-m MODIS imagery (2001–2010). Remote Sens. Environ. 2012, 126, 84–103.
Kiptala, J.; Mohamed, Y.; Mul, M.; Cheema, M.; der Zaag, P.V. Land use and land cover classification
using phenological variability from MODIS vegetation in the Upper Pangani River Basin, Eastern Africa.
Phys. Chem. Earth, Parts A/B/C 2013, 66, 112–122.
Leinenkugel, P.; Kuenzer, C.; Oppelt, N.; Dech, S. Characterisation of land surface phenology and land
cover based on moderate resolution satellite data in cloud prone areas—A novel product for the Mekong
Basin. Remote Sens. Environ. 2013, 136, 180–198.
Hansen, M.C.; Townshend, J.R.G.; DeFries, R.S.; Carroll, M. Estimation of tree cover using MODIS data
at global, continental and regional/local scales. Int. J. Remote Sens. 2005, 26, 4359–4380.
Hijmans, R.J.; Cameron, S.E.; Parra, J.L.; Jones, P.G.; Jarvis, A. Very high resolution interpolated climate
surfaces for global land areas. Int. J. Climatol. 2005, 25, 1965–1978.
Miao, C.; Ni, J.; Borthwick, A.G.; Yang, L. A preliminary estimate of human and natural contributions
to the changes in water discharge and sediment load in the Yellow River. Glob. Planet. Chang. 2011,
76, 196–205.
Chen, Y.; Syvitski, J.; Gao, S.; Overeem, I.; Kettner, A. Socio-economic impacts on flooding: A 4000-Year
history of the Yellow River, China. AMBIO 2012, 41, 682–698.
Chen, L.; Wei, W.; Fu, B.; Lu, Y. Soil and water conservation on the Loess Plateau in China: Review and
perspective. Prog. Phys. Geogr. 2007, 31, 389–403.
Olson, D.M.; Dinerstein, E.; Wikramanayake, E.D.; Burgess, N.D.; Powell, G.V.N.; Underwood, E.C.;
D’amico, J.A.; Itoua, I.; Strand, H.E.; Morrison, J.C.; et al. Terrestrial ecoregions of the world: A new map
of life on earth. BioScience 2001, 51, 933–938.
Wang, H.; Yang, Z.; Saito, Y.; Liu, J.P.; Sun, X.; Wang, Y. Stepwise decreases of the Huanghe (Yellow River)
sediment load (1950–2005): Impacts of climate change and human activities. Glob. Planet. Chang. 2007,
57, 331–354.
Ran, L.; Lu, X.; Xin, Z.; Yang, X. Cumulative sediment trapping by reservoirs in large river basins: A case
study of the Yellow River basin. Glob. Planet. Chang. 2013, 100, 308–319.
Huete, A.; Didan, K.; Miura, T.; Rodriguez, E.; Gao, X.; Ferreira, L. Overview of the radiometric and
biophysical performance of the MODIS vegetation indices. Remote Sens. Environ. 2002, 83, 195–213.
R Core Team. R: A Language and Environment for Statistical Computing; R Foundation for Statistical
Computing: Vienna, Austria, 2014.
Mattiuzzi, M.; Verbesselt, J.; Stevens, F.; Mosher, S.; Hengl, T.; Klisch, A.; Evans, B.; Lobo, A.
MODIS: MODIS Acquisition and Processing Package. Available online: http://R-Forge.R-project.org/
projects/modis/ (accessed on 10 February 2015).
Jönsson, P.; Eklundh, L. TIMESAT—A program for analyzing time-series of satellite sensor data.
Comput. Geosci. 2004, 30, 833–845.
Savitzky, A.; Golay, M.J.E. Smoothing and differentiation of data by simplified least squares procedures.
Anal. Chem. 1964, 36, 1627–1639.
Chang, J.; Hansen, M.C.; Pittman, K.; Carroll, M.; DiMiceli, C. Corn and soybean mapping in
the United States using MODIS time-series data sets. Agron. J. 2007, 99, 1654–1664.
Conrad, C.; Colditz, R.R.; Dech, S.; Klein, D.; Vlek, P.L.G. Temporal segmentation of MODIS time series for
improving crop classification in Central Asian irrigation systems. Int. J. Remote Sens. 2011, 32, 8763–8778.
Colditz, R.; Schmidt, M.; Conrad, C.; Hansen, M.; Dech, S. Land cover classification with coarse spatial
resolution data to derive continuous and discrete maps for complex regions. Remote Sens. Environ. 2011,
115, 3264–3275.
Congalton, R.; Green, K. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices;
Mapping Science, CRC Press: Boca Raton, FL, USA, 2009.
Congalton, R.G. A comparison of sampling schemes used in generating error matrices for assessing the
accuracy of maps generated from remotely sensed data. Photogram. Eng. Remote Sens. 1988, 54, 587–592.
Remote Sens. 2016, 8, 186
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
24 of 25
Jensen, J.R. Introductory Digital Image Processing: A Remote Sensing Perspective; Prentice Hall: Upper Saddle
River, NJ, USA, 2005.
Congalton, R.G. A review of assessing the accuracy of classifications of remotely sensed data.
Remote Sens. Environ. 1991, 37, 35–46.
Chan, J.C.W.; Paelinckx, D.
Evaluation of random forest and adaboost tree-based ensemble
classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery.
Remote Sens. Environ. 2008, 112, 2999–3011.
Lawrence, R.L.; Wood, S.D.; Sheley, R.L. Mapping invasive plants using hyperspectral imagery and
Breiman Cutler classifications (randomForest). Remote Sens. Environ. 2006, 100, 356–362.
Gislason, P.O.; Benediktsson, J.A.; Sveinsson, J.R. Random forests for land cover classification.
Pattern Recognit. Lett. 2006, 27, 294–300.
Breiman, L. Random Forests. Machine Learning; Springer: New York, NY, USA, 2001; pp. 5–32.
Liaw, A.; Wiener, M. Classification and regression by randomForest. R News 2002, 2, 18–22.
Kuhn, M.; Wing, J.; Weston, S.; Williams, A.; Keefer, C.; Engelhardt, A.; Cooper, T.; Mayer, Z.;
Kenkel, B.; the R Core Team, et al.
Caret: Classification and Regression Training. Available online:
http://CRAN.R-project. org/package=caret (accessed on 22 August 2015).
Kursa, M.B.; Rudnicki, W.R. Feature selection with the Boruta package. J. Stat. Softw. 2010, 36, 1–13.
Adelabu, S.; Mutanga, O.; Adam, E. Testing the reliability and stability of the internal accuracy assessment
of random forest for classifying tree defoliation levels using different validation methods. Geocarto Int.
2015, 30, 810–821.
Foody, G. Status of land cover classification accuracy assessment. Remote Sens. Environ. 2002, 80, 185–201.
NOAA-NGDC. Night-Time Lights of the World (Version 4). Available online: http://ngdc.noaa.gov/eog/
dmsp/downloadV4composites.html (accessed on 22 August 2015).
Small, C.; Pozzi, F.; Elvidge, C. Spatial analysis of global urban extent from DMSP-OLS night lights.
Remote Sens. Environ. 2005, 96, 277–291.
Li, L.; Friedl, M.A.; Xin, Q.; Gray, J.; Pan, Y.; Frolking, S. Mapping crop cycles in China using MODIS-EVI
time series. Remote Sens. 2014, 6, 2473–2493.
Liu, J.; Wu, P.; Wang, Y.; Zhao, X.; Sun, S.; Cao, X. Impacts of changing cropping pattern on virtual water
flows related to crops transfer: A case study for the Hetao irrigation district, China. J. Sci. Food Agric.
2014, 94, 2992–3000.
Guobin, L. Soil conservation and sustainable agriculture on the Loess Plateau: Challenges and prospects.
Ambio 1999, 28, 663–668.
Zhou, H.; Rompaey, A.V.; Wang, J. Detecting the impact of the “Grain for Green” program on the mean
annual vegetation cover in the Shaanxi province, China using SPOT-VGT NDVI data. Land Use Policy
2009, 26, 954–960.
Clark, M.L.; Aide, T.M.; Grau, H.R.; Riner, G. A scalable approach to mapping annual land cover
at 250 m using MODIS time series data: A case study in the Dry Chaco ecoregion of South America.
Remote Sens. Environ. 2010, 114, 2816–2832.
Blanco, P.D.; Colditz, R.R.; Saldaña, G.L.; Hardtke, L.A.; Llamas, R.M.; Mari, N.A.; Fischer, A.; Caride, C.;
Aceñolaza, P.G.; del Valle, H.F.; et al. A land cover map of Latin America and the Caribbean in the
framework of the SERENA project. Remote Sens. Environ. 2013, 132, 13–31.
Liu, J.; Li, Z.; Zhang, X.; Li, R.; Liu, X.; Zhang, H. Responses of vegetation cover to the Grain for
Green Program and their driving forces in the He-Long region of the middle reaches of the Yellow River.
J. Arid Land 2013, 5, 511–520.
Gao, P.; Niu, X.; Wang, B.; Zheng, Y. Land use changes and its driving forces in hilly ecological restoration
area based on gis and rs of northern china. Sci. Rep. 2015, 5, 1–5.
Zhai, J.; Liu, R.; Liu, J.; Huang, L.; Qin, Y. Human-induced Landcover changes drive a diminution of land
surface albedo in the Loess Plateau (China). Remote Sens. 2015, 7, 2926–2941.
Eisfelder, C.; Kuenzer, C. Investigating fourteen years of net primary productivity based on remote
sensing data for China. In Remote Sensing Time Series; Springer International Publishing: Cham,
Switzerland, 2015; pp. 269–288.
Feng, X.; Fu, B.; Lu, N.; Zeng, Y.; Wu, B. How ecological restoration alters ecosystem services: An analysis
of carbon sequestration in China’s Loess Plateau. Sci. Rep. 2013, 3, 1–5.
Remote Sens. 2016, 8, 186
72.
73.
74.
75.
76.
77.
78.
79.
80.
25 of 25
Ghose, B. Food security and food self-sufficiency in China: From past to 2050. Food Energy Secur. 2014,
3, 86–95.
Matson, P.A.; Parton, W.J.; Power, A.G.; Swift, M.J. Agricultural intensification and ecosystem properties.
Science 1997, 277, 504–509.
Tilman, D.; Fargione, J.; Wolff, B.; D’Antonio, C.; Dobson, A.; Howarth, R.; Schindler, D.;
Schlesinger, W.H.; Simberloff, D.; Swackhamer, D. Forecasting agriculturally driven global environmental
change. Science 2001, 292, 281–284.
Rudel, T.K.; Schneider, L.; Uriarte, M.; Turner, B.L.; DeFries, R.; Lawrence, D.; Geoghegan, J.; Hecht, S.;
Ickowitz, A.; Lambin, E.F.; et al. Agricultural intensification and changes in cultivated areas, 1970–2005.
Proc. Natl. Acad. Sci. USA 2009, 106, 20675–20680.
Siebert, S.; Portmann, F.T.; Döll, P. Global patterns of cropland use intensity. Remote Sens. 2010,
2, 1625–1643.
Wang, J.; Liu, Y.; Liu, Z. Spatio-temporal patterns of cropland conversion in response to the “Grain for
Green Project” in China’s loess hilly region of Yanchuan County. Remote Sens. 2013, 5, 5642–5661.
Meng, E.C.; Hu, R.; Shi, X.; Zhang, S. Maize in China: Production Systems, Constraints, and Research Priorities;
Maize Production Systems Papers 7648; (CIMMYT) International Maize and Wheat Improvement Center:
El Batan, Mexico, 2006.
Fernandes, R.; Fraser, R.; Latifovic, R.; Cihlar, J.; Beaubien, J.; Du, Y. Approaches to fractional
land cover and continuous field mapping: A comparative assessment over the BOREAS study region.
Remote Sens. Environ. 2004, 89, 234–251.
Gessner, U.; Machwitz, M.; Conrad, C.; Dech, S. Estimating the fractional cover of growth forms
and bare surface in savannas. A multi-resolution approach based on regression tree ensembles.
Remote Sens. Environ. 2013, 129, 90–102.
c 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open
access article distributed under the terms and conditions of the Creative Commons by
Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).